Method and system for making binary predictions for a subject using historical data obtained from multiple subjects

Abstract

A method for making binary predictions for a subject involves obtaining historical data for multiple subjects, the historical data including, for each subject, a feature set and a binary outcome, generating training data from the historical data, and training a predictive model using the training data to predict the outcomes based on the feature sets. The method further includes obtaining historical data including a feature set for a subject under consideration, and predicting a binary outcome for the subject under consideration, based on the feature set associated with the subject under consideration.

What is claimed is: 
     
       1. A method for making binary predictions for a subject, the method comprising:
 obtaining historical data for a plurality of subjects, the historical data comprising, for each subject, a feature set and a binary outcome, wherein the binary outcome is one of a first outcome and a second outcome, wherein the plurality of subjects have an initial ratio of the first outcome to the second outcome, the initial ratio having a bias; 
 sampling the historical data of the plurality of subjects according to the binary outcome to create training data having a target ratio of the first outcome to the second outcome, the target ratio correcting the bias; 
 training, through a plurality of stages, a predictive model using the training data to predict the outcomes based on the feature sets, wherein training the predictive model comprises, for each of the plurality of stages: 
 executing a loss function on an existing set of decision trees to determine a loss denoting a predictive accuracy of the existing set of decision trees, and 
 adding a decision tree to the existing set of decision trees, wherein adding the decision tree comprises using a gradient descent procedure to parameterize the decision tree according to the loss; 
 obtaining historical data comprising a feature set for a subject under consideration; and 
 predicting, after training and by the predictive model, a binary outcome for the subject under consideration, based on the feature set associated with the subject under consideration. 
 
     
     
       2. The method of  claim 1 , wherein the subject under consideration is a business, and
 wherein the prediction of the binary outcome for the subject under consideration comprises predicting one selected from a group consisting of success and failure of the business. 
 
     
     
       3. The method of  claim 2 , wherein the feature sets comprise financial attributes. 
     
     
       4. The method of  claim 1 , further comprising, prior to training the predictive model, enhancing each of the feature sets with an engineered feature. 
     
     
       5. The method of  claim 4 , wherein the engineered feature is a ratio of features in the feature set. 
     
     
       6. The method of  claim 1 , wherein sampling the historical data comprises a pairwise sampling that selects, for a subject with a first feature and a positive outcome, a subject with a second feature matching the first feature and a negative outcome, to be included in the training data. 
     
     
       7. The method of  claim 1 , wherein one of the binary outcomes in the historical data is not initially known, and wherein obtaining the historical data comprises deriving the initially unknown outcome from cues in the historical data. 
     
     
       8. The method of  claim 1 , wherein the predictive model is an xgboost model. 
     
     
       9. A system for making binary predictions for a subject, the system comprising:
 a subject database comprising historical data for a plurality of subjects; 
 a computer processor; 
 a predictive model training engine executing on the computer processor configured to: 
 obtain the historical data for the plurality of subjects, the historical data comprising, for each subject, a feature set and a binary outcome, wherein the binary outcome is one of a first outcome and a second outcome, wherein the plurality of subjects have an initial ratio of the first outcome to the second outcome, the initial ratio having a bias; 
 sampling the historical data of the plurality of subjects according to the binary outcome to create training data having a target ratio of the first outcome to the second outcome, the target ratio correcting the bias; 
 generate training data from the obtained historical data; and 
 train, through a plurality of stages, a predictive model using the training data to predict the outcomes based on the feature sets, wherein training the predictive model comprises, for each of the plurality of stages: 
 executing a loss function on an existing set of decision trees to determine a loss denoting a predictive accuracy of the existing set of decision trees, and adding a decision tree to the existing set of decision trees, wherein adding the decision tree comprises using a gradient descent procedure to parameterize the decision tree according to the loss; 
 a prediction engine executing on the computer processor configured to: 
 obtain the historical data comprising a feature set for a subject under consideration; and 
 predict, after training and by the predictive model, a binary outcome for the subject under consideration, based on the feature set associated with the subject under consideration. 
 
     
     
       10. The system of  claim 9 , wherein the predictive model training engine is further configured to, prior to training the predictive model, enhance each of the feature sets with an engineered feature. 
     
     
       11. The system of  claim 9 , wherein sampling the historical data comprises a pairwise sampling that selects, for a subject with a first feature and a positive outcome, a subject with a second feature matching the first feature and a negative outcome, to be included in the training data. 
     
     
       12. The system of  claim 9 ,
 wherein one of the binary outcomes in the historical data is not initially known, and 
 wherein obtaining the historical data comprises deriving the initially unknown outcome from cues in the historical data. 
 
     
     
       13. The system of  claim 9 , wherein the predictive model is an xgboost model. 
     
     
       14. A non-transitory computer readable medium comprising computer readable program code for causing a computer system to:
 obtain historical data for a plurality of subjects, the historical data comprising, for each subject, a feature set and a binary outcome, wherein the binary outcome is one of a first outcome and a second outcome, wherein the plurality of subjects have an initial ratio of the first outcome to the second outcome, the initial ratio having a bias; 
 sampling the historical data of the plurality of subjects according to the binary outcome to create training data having a target ratio of the first outcome to the second outcome, the target ratio correcting the bias; 
 train, through a plurality of stages, a predictive model using the training data to predict the outcomes based on the feature sets, wherein training the predictive model comprises, for each of the plurality of stages: 
 executing a loss function on an existing set of decision trees to determine a loss denoting a predictive accuracy of the existing set of decision trees, and 
 adding a decision tree to the existing set of decision trees, wherein adding the decision tree comprises using a gradient descent procedure to parameterize the decision tree according to the loss; 
 obtain historical data comprising a feature set for a subject under consideration; and 
 predict, after training and by the predictive model, a binary outcome for the subject under consideration, based on the feature set associated with the subject under consideration. 
 
     
     
       15. The non-transitory computer readable medium of  claim 14 , further comprising computer readable program code for causing the computer system to, prior to training the predictive model, enhancing each of the feature sets with an engineered feature. 
     
     
       16. The non-transitory computer readable medium of  claim 14 , wherein sampling the historical data comprises a pairwise sampling that selects, for a subject with a first feature and a positive outcome, a subject with a second feature matching the first feature and a negative outcome, to be included in the training data. 
     
     
       17. The non-transitory computer readable medium of  claim 14 ,
 wherein one of the binary outcomes in the historical data is not initially known, and 
 wherein obtaining the historical data comprises deriving the initially unknown outcome from cues in the historical data.